this fails because pandas changed the way they treat dask arrays passed to their constructor: we now get a series of object dtype filled with 0d dask arrays.

We can work around that by computing x before passing it to pandas.DataFrame in the to_dask_dataframe tests: python expected_pd = pd.DataFrame({"a": x.compute(), "b": y}, index=pd.Index(t, name="t")) or by assigning the initial numpy array to x and converting it to a dask array immediately before passing it to the xr.Dataset constructor: ```python x = np.random.randn(10)

...

ds = xr.Dataset({"a": ("t", da.from_array(x, chunks=4)), "b": ("t", y), "t": ("t", t)}) ``` (and something similar for the other two failing tests)

But maybe we should just ask the pandas devs whether that was intentional?